Source
- URL: https://x.com/rianflo/status/2001080321802977727
- Author: Florian Hoenig (@rianflo)
- Posted: 2025-12-17 00:01:41
Thread
1/ @rianflo
Significant amount of bloat dropped :-) 🎉🥳🎈
I’d move it all to the GPU though. A single kernel that launches your pipeline steps per frame or per whatever. Pretty sure NV can do it. AMD I have no deep knowledge. @NOTimothyLottes ?
2/ @NOTimothyLottes
@rianflo
From my perspective being compute-only for graphics, few reasons to launch kernels because I’d always use a fixed amount of VGPRs. The larger problem is compilers that cannot keep up with the large program sizes (in compile time or register allocation).
Branches: 2025-12-17-rianflo-question-is-how-much-overhead-of-course-even-not
3/ @NOTimothyLottes
@rianflo
Huge amount of pitfalls with compute-side dynamic dispatch of work (NV had it back when I worked there BTW), from long latency chains, to overheads tracking work completion, preemption, state restore, etc. It’s not the golden answer people think it is. It is a slow path.
4/ @NOTimothyLottes
@rianflo
There are however a bunch of things we could trivially do GPU side that are currently CPU side, like computing best tiling modes for images, computing image {size, alignment, etc}, building descriptors, etc. Lots of other driver stuff could be done on the GPU for sure.
5/ @NOTimothyLottes
@rianflo
My plan for 2026 is actually to build my open source AMD SteamOS/Linux driver for a future compute-graphics API, and it will be doing some radical stuff since I’m also doing shader binary code generation too :) Hopefully you all get live examples.
6/ @bkaradzic
@NOTimothyLottes @rianflo
Can’t wait! This compute-graphics API is something like compute + meshlets + fragment shaders only graphics API?
7/ @NOTimothyLottes
@bkaradzic @rianflo
Compute as in only compute shader dispatches, no fixed function graphics (no fragment/etc). And graphics via pure compute.
8/ @Nicolas_Lopez_
@NOTimothyLottes @bkaradzic @rianflo
Considering how much we already software rasterize, we are not so far from it :)
9/ @rianflo
@Nicolas_Lopez_ @NOTimothyLottes @bkaradzic
yeah but then someone insists on a couple of large triangles and wants to access the rasterizer hw again, which is supposedly too small of die area to be even worth ripping it out.